[PowerPC] Initial support for the VSX instruction set
VSX is an ISA extension supported on the POWER7 and later cores that enhances
floating-point vector and scalar capabilities. Among other things, this adds
<2 x double> support and generally helps to reduce register pressure.
The interesting part of this ISA feature is the register configuration: there
are 64 new 128-bit vector registers, the 32 of which are super-registers of the
existing 32 scalar floating-point registers, and the second 32 of which overlap
with the 32 Altivec vector registers. This makes things like vector insertion
and extraction tricky: this can be free but only if we force a restriction to
the right register subclass when needed. A new "minipass" PPCVSXCopy takes care
of this (although it could do a more-optimal job of it; see the comment about
unnecessary copies below).
Please note that, currently, VSX is not enabled by default when targeting
anything because it is not yet ready for that. The assembler and disassembler
are fully implemented and tested. However:
- CodeGen support causes miscompiles; test-suite runtime failures:
MultiSource/Benchmarks/FreeBench/distray/distray
MultiSource/Benchmarks/McCat/08-main/main
MultiSource/Benchmarks/Olden/voronoi/voronoi
MultiSource/Benchmarks/mafft/pairlocalalign
MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4
SingleSource/Benchmarks/CoyoteBench/almabench
SingleSource/Benchmarks/Misc/matmul_f64_4x4
- The lowering currently falls back to using Altivec instructions far more
than it should. Worse, there are some things that are scalarized through the
stack that shouldn't be.
- A lot of unnecessary copies make it past the optimizers, and this needs to
be fixed.
- Many more regression tests are needed.
Normally, I'd fix these things prior to committing, but there are some
students and other contributors who would like to work this, and so it makes
sense to move this development process upstream where it can be subject to the
regular code-review procedures.
llvm-svn: 203768
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 680112d..a35c83c 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -505,7 +505,7 @@
setOperationAction(ISD::MUL, MVT::v4f32, Legal);
setOperationAction(ISD::FMA, MVT::v4f32, Legal);
- if (TM.Options.UnsafeFPMath) {
+ if (TM.Options.UnsafeFPMath || Subtarget->hasVSX()) {
setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
}
@@ -532,6 +532,40 @@
setCondCodeAction(ISD::SETO, MVT::v4f32, Expand);
setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand);
+
+ if (Subtarget->hasVSX()) {
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
+ setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
+
+ setOperationAction(ISD::MUL, MVT::v2f64, Legal);
+ setOperationAction(ISD::FMA, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
+ setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
+
+ // Share the Altivec comparison restrictions.
+ setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETUGT, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETUGE, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETULT, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETULE, MVT::v2f64, Expand);
+
+ setCondCodeAction(ISD::SETO, MVT::v2f64, Expand);
+ setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand);
+
+ addRegisterClass(MVT::f64, &PPC::VSRCRegClass);
+
+ addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass);
+ addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass);
+ }
}
if (Subtarget->has64BitSupport()) {
@@ -2094,6 +2128,7 @@
case MVT::v8i16:
case MVT::v4i32:
case MVT::v4f32:
+ case MVT::v2f64:
RC = &PPC::VRRCRegClass;
break;
}
@@ -2340,7 +2375,8 @@
// Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
- ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
+ ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8 ||
+ ObjectVT==MVT::v2f64) {
if (isVarArg) {
MinReservedArea = ((MinReservedArea+15)/16)*16;
MinReservedArea += CalculateStackSlotSize(ObjectVT,
@@ -2497,6 +2533,7 @@
case MVT::v4i32:
case MVT::v8i16:
case MVT::v16i8:
+ case MVT::v2f64:
// Note that vector arguments in registers don't reserve stack space,
// except in varargs functions.
if (VR_idx != Num_VR_Regs) {
@@ -2959,7 +2996,8 @@
EVT ArgVT = Outs[i].VT;
// Varargs Altivec parameters are padded to a 16 byte boundary.
if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
- ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
+ ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8 ||
+ ArgVT==MVT::v2f64) {
if (!isVarArg && !isPPC64) {
// Non-varargs Altivec parameters go after all the non-Altivec
// parameters; handle those later so we know how much padding we need.
@@ -4143,6 +4181,7 @@
case MVT::v4i32:
case MVT::v8i16:
case MVT::v16i8:
+ case MVT::v2f64:
if (isVarArg) {
// These go aligned on the stack, or in the corresponding R registers
// when within range. The Darwin PPC ABI doc claims they also go in
@@ -6917,7 +6956,8 @@
if ((VT == MVT::f32 && PPCSubTarget.hasFRES()) ||
(VT == MVT::f64 && PPCSubTarget.hasFRE()) ||
- (VT == MVT::v4f32 && PPCSubTarget.hasAltivec())) {
+ (VT == MVT::v4f32 && PPCSubTarget.hasAltivec()) ||
+ (VT == MVT::v2f64 && PPCSubTarget.hasVSX())) {
// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
// For the reciprocal, we need to find the zero of the function:
@@ -6979,7 +7019,8 @@
if ((VT == MVT::f32 && PPCSubTarget.hasFRSQRTES()) ||
(VT == MVT::f64 && PPCSubTarget.hasFRSQRTE()) ||
- (VT == MVT::v4f32 && PPCSubTarget.hasAltivec())) {
+ (VT == MVT::v4f32 && PPCSubTarget.hasAltivec()) ||
+ (VT == MVT::v2f64 && PPCSubTarget.hasVSX())) {
// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
// For the reciprocal sqrt, we need to find the zero of the function:
@@ -7891,6 +7932,7 @@
unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
if (ISD::isNON_EXTLoad(N) && VT.isVector() &&
TM.getSubtarget<PPCSubtarget>().hasAltivec() &&
+ // FIXME: Update this for VSX!
(VT == MVT::v16i8 || VT == MVT::v8i16 ||
VT == MVT::v4i32 || VT == MVT::v4f32) &&
LD->getAlignment() < ABIAlignment) {
@@ -8314,6 +8356,9 @@
}
} else if (Constraint == "wc") { // individual CR bits.
return C_RegisterClass;
+ } else if (Constraint == "wa" || Constraint == "wd" ||
+ Constraint == "wf" || Constraint == "ws") {
+ return C_RegisterClass; // VSX registers.
}
return TargetLowering::getConstraintType(Constraint);
}
@@ -8335,6 +8380,13 @@
// Look at the constraint type.
if (StringRef(constraint) == "wc" && type->isIntegerTy(1))
return CW_Register; // an individual CR bit.
+ else if ((StringRef(constraint) == "wa" ||
+ StringRef(constraint) == "wd" ||
+ StringRef(constraint) == "wf") &&
+ type->isVectorTy())
+ return CW_Register;
+ else if (StringRef(constraint) == "ws" && type->isDoubleTy())
+ return CW_Register;
switch (*constraint) {
default:
@@ -8393,6 +8445,9 @@
}
} else if (Constraint == "wc") { // an individual CR bit.
return std::make_pair(0U, &PPC::CRBITRCRegClass);
+ } else if (Constraint == "wa" || Constraint == "wd" ||
+ Constraint == "wf" || Constraint == "ws") {
+ return std::make_pair(0U, &PPC::VSRCRegClass);
}
std::pair<unsigned, const TargetRegisterClass*> R =